mates of the returns to attendance obtained from specifications that include
appropriate control variables are still likely to be biased and inconsistent,
to the extent that they incorrectly attribute to attendance the effect of the
component of ability and motivation not captured by the controls.
One possible solution would be to find appropriate instruments for atten-
dance. However, it is generally quite difficult to find variables correlated with
attendance but uncorrelated with unobservable ability, effort and motivation.
An alternative route, followed in this paper, is to exploit the variability of
attendance and performance in the time dimension, if a panel data set is
available. This allows to take into account time-invariant unobservable fac-
tors that affect both attendance and performance, and therefore to eliminate
the omitted variable bias that characterizes estimates of the effect of atten-
dance on performance based on cross-sectional data.
For the analysis presented in this study, we collected observations on the
performance of 766 Introductory Microeconomics students on several tests,
and their attendance levels at lectures and classes covering the material ex-
amined on those tests. We also have information on proxies for ability (high
school grade, grade point average, exam speed, and proficiency in calculus),
effort (number of study hours) and motivation (subject and teacher evalua-
tion), candidate instruments for attendance and a number of other individual
characteristics. We can therefore compare the results obtained with three ap-
proaches: OLS controlling for unobservable factors with proxy variables; in-
strumental variables (2SLS) for attendance; panel estimators (random effects
and fixed effects).
We find that both OLS and IV estimates of the effects of attendance on
performance are positive and significant. However, neither proxy variables
nor instrumental variables provide a viable solution to the omitted variable
bias: proxy variables do not capture all the correlation between the regressor
of interest and the omitted factors, while candidate instrumental variables
are found to be correlated with the error term. When we eliminate the
omitted variable bias, using a fixed effect estimator, the point estimate for
attendance is about half the size of the OLS and IV estimates, but the effect
on performance remains positive and significant. We also find that lecture
and classes have a similar effect on performance individually, although their
impact cannot be identified separately. Overall, the results indicate that
teaching is a key factor for student learning.
The remainder of the paper is structured as follows. Section 2 reviews
the empirical literature on student attendance and academic performance.